Engineering Science and Technology, an International Journal (Jun 2024)
Machine-learning-based precise cost-efficient NO2 sensor calibration by means of time series matching and global data pre-processing
Abstract
Air pollution remains a considerable contemporary challenge affecting life quality, the environment, and economic well-being. It encompasses an array of pollutants—gases, particulate matter, biological molecules—emanating from sources such as vehicle emissions, industrial activities, agriculture, and natural occurrences. Nitrogen dioxide (NO2), a harmful gas, is particularly abundant in densely populated urban areas. Given its detrimental impact on health and the environment, precise monitoring of NO2 levels is crucial for devising effective strategies to mitigate risks. However, precise measurement of NO2 presents challenges as it traditionally relies on expensive and heavy (therefore, stationary) equipment. This has led to the pursuit of more affordable alternatives, though their dependability is frequently questionable. This study introduces an innovative technique for precise calibration of low-cost NO2 sensors. Our methodology involves statistical preprocessing of sensor measurements to align their distributions with reference data. The core of the calibration model is an artificial neural network (ANN), trained to synchronize sensor and reference time series measurements. It incorporates environmental variables such as temperature, humidity, and atmospheric pressure, along with readings from redundant NO2 sensors for cross-referencing, and short time series of primary sensor NO2 measurements. This enables efficient learning of typical sensor changes over time in relation to these factors. Additionally, an interpolative kriging model serves as an auxiliary surrogate to enhance the correction process’s reliability. Validation using an autonomous monitoring platform from Gdansk University of Technology, Poland, and public reference station data gathered over five months shows remarkable calibration accuracy, with a correlation coefficient close to 0.95 and RMSE of 2.4 µg/m3. These results position the corrected sensor as an attractive and cost-effective alternative to conventional NO2 measurement methods.
